generic: 6.12: backport bcm63xx-spi CS deassertion fix
authorÁlvaro Fernández Rojas <[email protected]>
Sun, 30 Nov 2025 16:09:46 +0000 (17:09 +0100)
committerÁlvaro Fernández Rojas <[email protected]>
Mon, 1 Dec 2025 08:12:36 +0000 (09:12 +0100)
Backport fix for premature SPI CS deassertion on RX-only transactions from
linux v6.18, which is needed for bmips (bcm63xx).

Signed-off-by: Álvaro Fernández Rojas <[email protected]>
target/linux/generic/backport-6.12/830-v6.18-spi-bcm63xx-fix-premature-CS-deassertion-on-RX-only-transactions.patch [new file with mode: 0644]

diff --git a/target/linux/generic/backport-6.12/830-v6.18-spi-bcm63xx-fix-premature-CS-deassertion-on-RX-only-transactions.patch b/target/linux/generic/backport-6.12/830-v6.18-spi-bcm63xx-fix-premature-CS-deassertion-on-RX-only-transactions.patch
new file mode 100644 (file)
index 0000000..6a8889e
--- /dev/null
@@ -0,0 +1,59 @@
+From fd9862f726aedbc2f29a29916cabed7bcf5cadb6 Mon Sep 17 00:00:00 2001
+From: Hang Zhou <[email protected]>
+Date: Mon, 17 Nov 2025 01:08:35 +1100
+Subject: [PATCH] spi: bcm63xx: fix premature CS deassertion on RX-only
+ transactions
+
+On BCM6358 (and also observed on BCM6368) the controller appears to
+only generate as many SPI clocks as bytes that have been written into
+the TX FIFO. For RX-only transfers the driver programs the transfer
+length in SPI_MSG_CTL but does not write anything into the FIFO, so
+chip select is deasserted early and the RX transfer segment is never
+fully clocked in.
+
+A concrete failing case is a three-transfer MAC address read from
+SPI-NOR:
+  - TX 0x03 (read command)
+  - TX 3-byte address
+  - RX 6 bytes (MAC)
+
+In contrast, a two-transfer JEDEC-ID read (0x9f + 6-byte RX) works
+because the driver uses prepend_len and writes dummy bytes into the
+TX FIFO for the RX part.
+
+Fix this by writing 0xff dummy bytes into the TX FIFO for RX-only
+segments so that the number of bytes written to the FIFO matches the
+total message length seen by the controller.
+
+Fixes: b17de076062a ("spi/bcm63xx: work around inability to keep CS up")
+
+Signed-off-by: Hang Zhou <[email protected]>
+Link: https://patch.msgid.link/[email protected]
+Signed-off-by: Mark Brown <[email protected]>
+---
+ drivers/spi/spi-bcm63xx.c | 14 ++++++++++++++
+ 1 file changed, 14 insertions(+)
+
+--- a/drivers/spi/spi-bcm63xx.c
++++ b/drivers/spi/spi-bcm63xx.c
+@@ -247,6 +247,20 @@ static int bcm63xx_txrx_bufs(struct spi_
+               if (t->rx_buf) {
+                       do_rx = true;
++
++                      /*
++                       * In certain hardware implementations, there appears to be a
++                       * hidden accumulator that tracks the number of bytes written into
++                       * the hardware FIFO, and this accumulator overrides the length in
++                       * the SPI_MSG_CTL register.
++                       *
++                       * Therefore, for read-only transfers, we need to write some dummy
++                       * value into the FIFO to keep the accumulator tracking the correct
++                       * length.
++                       */
++                      if (!t->tx_buf)
++                              memset_io(bs->tx_io + len, 0xFF, t->len);
++
+                       /* prepend is half-duplex write only */
+                       if (t == first)
+                               prepend_len = 0;